Ask an assistant for the best cordless vacuum and it does not answer from memory. It pulls somewhere between nine and nineteen pages, weighs them, and writes a recommendation. A study published in September 2026 by GetCited and DataPulse Research opened those pages up and read the small print. Roughly one in four had been written under an affiliate, partner or advertising arrangement, and for some product categories the figure was far higher. For a publisher that earns money from product recommendations, that is both reassuring and uncomfortable: the work is being used, but the reader no longer needs to visit to use it.
What did the GetCited study actually measure?
The study put 3,334 English best-product and comparison questions to ChatGPT, Google AI Overview and Perplexity, spanning 748 product and service types across 15 sectors. A matching German set brought the total to 6,668 questions. From the matched English answers it collected 106,758 citations pointing at 139,033 unique URLs, then scanned the pages with a headless browser that renders JavaScript, checking 22 commercial markers in three layers: affiliate disclosures, partner-content notices and advertising labels. The data was collected in April 2026.
The headline result, as reported by PPC Land, is that 25.4 per cent of English citations led to a page with a disclosed commercial interest, and that 86 per cent of answers contained at least one such source. The bulk of it was affiliate disclosure, at 21.4 per cent of citations. Partner content accounted for 2.9 per cent, dedicated commerce sections such as Forbes Advisor and Wirecutter for 2.0 per cent, and advertising labels for 1.8 per cent. Those layers overlap, so they do not add up neatly.
Which AI assistant cites the most commercial content?
Perplexity leaned on commercial pages most heavily, at 31.1 per cent of its citations, against 23.6 per cent for Google AI Overview and 22.5 per cent for ChatGPT. The answer-level view tells a different story. ChatGPT included at least one commercial source in 91 per cent of answers, Perplexity in 89 per cent and Google in 77 per cent, partly because ChatGPT cites far more pages per answer, about 19 against roughly 10 for Perplexity and 9 for Google.
The share of answers built mostly from commercial sources was lower: 17 per cent for Perplexity, 10 per cent for Google and 5 per cent for ChatGPT. A larger source pool dilutes the commercial share per answer, but it also makes it almost certain that a commercial page appears somewhere in the mix.
Which product categories lean hardest on commercial sources?
Consumer electronics topped the sector table at 41.3 per cent of citations, followed by fitness and sports at 37.5 per cent, telecoms, VPN and hosting at 32.8 per cent, media and entertainment at 31.0 per cent, and baby, kids and pets at 30.1 per cent. Cordless vacuums reached 61 per cent across 120 citations, and fitness trackers 60 per cent across 121.
At the other end, travel and hospitality sat at 16.5 per cent, legal and professional services at 17.8 per cent, and finance and insurance at 19.9 per cent. The authors caution that low readings in areas such as clinics and legal services do not mean neutral sourcing. Those answers draw largely on provider pages and directories, which carry no disclosure a scanner can read.
Which publishers supply the commercial layer?
The supply is concentrated. The ten most-cited commercial sources were TechRadar (1,225 citations), Forbes Advisor (1,134), PCMag (964), CNET (616), Tom's Guide (599), RTINGS (573), NYT Wirecutter (472), Good Housekeeping (426), Healthline (411) and Wired (366). Together they supplied roughly a quarter of all commercially disclosed English citations.
That concentration matters for two reasons. It shows that assistants do not sample the open web evenly when the question is a purchase: they return to a short list of established review brands with testing methods, structured comparisons and a track record. And it means the long tail of smaller review sites is largely absent from the answers that shape buying decisions, however good its individual pages may be.
Why does the cited layer not pay for itself?
The interface that draws on these pages also removes the click that funds them. Future plc reported in May 2026 that organic eCommerce affiliate revenue for the six months to 31 March fell 24 per cent to 32.3 million pounds, while Google Search and Discover audiences each fell by about 20 per cent. TechRadar and Tom's Guide sit inside that group, and both appear in the GetCited top ten.
The pattern is not confined to affiliate. Ziff Davis took a $54.8 million goodwill impairment on its Health and Wellness segment in the second quarter of 2026, the vertical that independent measurement has repeatedly flagged as most exposed to AI answers. That charge was in health media rather than the product-review brands, so it is evidence of the wider squeeze and not of affiliate loss specifically.
The mechanics are the same in both cases. A citation is a credit line. It is not a visit, a pageview, a tracked affiliate click or a paid impression. The assistant reads the review, shortlists the product and often completes the recommendation in the conversation, so the commission trail that made the review worth writing is never triggered. The agentic-shopping shift makes this sharper, because agents can transact without any publisher link at all.
What should we not conclude from the study?
Four limits are worth stating plainly. First, the study is a floor. It counts disclosed commercial interest, and it cannot measure pages with undisclosed interests, so the true commercial share could be higher by an amount nobody knows. Second, it was run by an interested party: GetCited sells placement of comparison articles on established publishers, with pricing reported at $3,000 a month on a 12-month strategy. That does not make the numbers wrong, but it is context a reader should hold. Third, the data is from April 2026 and citation sets rotate quickly, so the shares describe a moment. Fourth, the study measures what gets cited, not what decides the recommendation.
The authors' own summary is fair on the last point: the recommendation is judged on the source's authority, not its commercial model. An affiliate disclosure does not stop a page being trusted. It tells you the page is part of an economy that the assistant is now sitting on top of.
What does this mean for publishers with commercial content?
The finding is more useful as a map of where value is being extracted than as a warning about bias. If a quarter of the pages an assistant relies on for a purchase answer are commercial content from a small set of outlets, then those outlets are supplying the raw material for a recommendation product they do not run and are not paid for.
Three practical steps follow. Measure the reads, not just the referrals: server-side logs show which assistants fetch your review pages, when, and how often, which is the only way to see demand that never arrives as a session. Treat structure as a supply-side asset, since the pages the study shows being cited are the ones with clear testing, comparison tables and named products. And decide deliberately what to allow, because the pages that earn the citation are the ones an assistant can read most cleanly.
blankspace works at that last point, at the CDN edge, where a live retrieval can be identified as it happens and recorded against the publisher's own logs. It is not a substitute for building pages worth citing, and it does not change what an assistant decides to recommend. It addresses the narrower question of what the publisher can do at the moment its page is read, given that the read is now where the value is created.
What does this mean for advertisers and brands?
For brands the study is an uncomfortable confirmation that presence in AI answers runs through third parties. The pages doing the persuading are mostly not brand-owned. They are reviews on publisher sites, which is why the market for placement on them is forming quickly and why disclosure will matter more, not less. A brand that pays to be featured on a review page is buying into the source an assistant then repeats.
The durable route for a brand is the one the study's own conclusion points to: be in the pages that assistants already treat as authoritative, on terms the publisher and the reader can see. Buying placement in a hidden layer is a short-term play in a system where the sources are checkable in one click.
Frequently asked questions
Do AI assistants cite affiliate content?
Yes. In the GetCited study of best-product questions, 21.4 per cent of citations led to pages carrying an affiliate disclosure, and 86 per cent of answers cited at least one page with a disclosed commercial interest. Perplexity, Google AI Overview and ChatGPT all did it, at differing rates.
Do AI assistants prefer commercial sources over independent ones?
The study does not show a preference, because it measured what was cited and not what was chosen over alternatives. Commercial pages made up 25.4 per cent of English citations, which means about three in four citations went to pages without a disclosed commercial interest. The finding is that commercial content is heavily represented, not that it is favoured.
Which publishers are cited most for best-product questions?
TechRadar, Forbes Advisor and PCMag led the commercially disclosed citations in the study, followed by CNET, Tom's Guide, RTINGS, NYT Wirecutter, Good Housekeeping, Healthline and Wired. Those ten supplied roughly a quarter of the commercially disclosed English citations in the dataset.
Does being cited by an AI assistant earn affiliate revenue?
Not by itself. A citation is a reference in the answer, and commission depends on a tracked click and a completed purchase. When the assistant reads the review and gives the recommendation without the reader visiting, no click occurs. Future plc's 24 per cent fall in organic eCommerce affiliate revenue over the six months to March 2026 is consistent with that pressure.
Can publishers see when AI assistants read their review pages?
Not in client-side analytics, because many AI retrievals never execute the JavaScript those tools rely on. Server-side or edge logs record the request itself, including the user agent and timing, which makes them the reliable way to see how often assistants fetch a given review page.